-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathTFNH.py
224 lines (184 loc) · 11.9 KB
/
TFNH.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
from setting import *
from ops import *
import scipy.io as sio
import numpy as np
import tensorflow as tf
import tensorflow.contrib.slim as slim
#import matplotlib.pyplot as plt
from calc_hammingranking import calc_map
class TFNH(object):
def __init__(self, sess):
self.I_tr = I_tr
self.T_tr = T_tr
self.L_tr = L_tr
self.I_te = I_te
self.T_te = T_te
self.L_te = L_te
self.pair = pair
self.pair_batch_size = pair_batch_size
self.img_batch_size = img_batch_size
self.txt_batch_size = txt_batch_size
self.BATCH_NUM = BATCH_NUM
self.batch_size = batch_size
self.TOTAL_EPOCH = TOTAL_EPOCH
self.all_num = all_num
self.hid_dim = hid_dim
self.hash_dim = hash_dim
self.dis_dim = dis_dim
self.dis_out_dim = dis_out_dim
self.img_dim = img_dim
self.txt_dim = txt_dim
self.lab_dim = lab_dim
self.fusion_dim = fusion_dim
self.lr_fn = lr_fn
self.lr_dn = lr_dn
self.kp = kp
self.fusion_net = fusion_net
self.classification_net = classification_net
self.discriminative_net1 = discriminative_net1
self.discriminative_net2 = discriminative_net2
self.build_model()
self.sess = sess
def build_model(self):
self.ph = {}
self.ph['fusion_input1'] = tf.placeholder(tf.float32, [None, self.fusion_dim])
self.ph['fusion_input2'] = tf.placeholder(tf.float32, [None, self.fusion_dim])
self.ph['fusion_input3'] = tf.placeholder(tf.float32, [None, self.fusion_dim])
self.ph['lab1'] = tf.placeholder(tf.float32, [None, self.lab_dim])
self.ph['lab2'] = tf.placeholder(tf.float32, [None, self.lab_dim])
self.ph['lab3'] = tf.placeholder(tf.float32, [None, self.lab_dim])
self.ph['kp'] = tf.placeholder(tf.float32)
# fusion network
self.f_feat = self.fusion_net(self.ph['fusion_input1'], self.hid_dim, self.hash_dim, self.ph['kp'])
self.i_feat = self.fusion_net(self.ph['fusion_input2'], self.hid_dim, self.hash_dim, self.ph['kp'], reuse=True)
self.t_feat = self.fusion_net(self.ph['fusion_input3'], self.hid_dim, self.hash_dim, self.ph['kp'], reuse=True)
# classification network
self.class_net1 = self.classification_net(self.f_feat, self.hash_dim, self.lab_dim)
self.class_net2 = self.classification_net(self.i_feat, self.hash_dim, self.lab_dim, reuse=True)
self.class_net3 = self.classification_net(self.t_feat, self.hash_dim, self.lab_dim, reuse=True)
# feature
self.dis_feat1 = tf.concat((self.f_feat, self.i_feat), 0)
self.dis_feat2 = tf.concat((self.f_feat, self.t_feat), 0)
# discriminator
self.dis_net1 = self.discriminative_net1(self.dis_feat1, self.dis_dim, self.dis_out_dim)
self.dis_net2 = self.discriminative_net2(self.dis_feat2, self.dis_dim, self.dis_out_dim)
# gragh loss
W1 = tf.matmul(self.ph['lab1'], tf.transpose(self.ph['lab1']))
D1 = tf.matrix_diag(tf.reduce_sum(W1, 1))
L1 = tf.add(W1, -D1)
self.graph_loss1 = beta[0] * tf.trace(tf.matmul(tf.matmul(tf.transpose(self.f_feat), L1), self.f_feat)) / (self.pair_batch_size * self.pair_batch_size)
d11 = tf.reduce_sum(tf.square(self.ph['fusion_input2']), 1)
d12 = tf.matmul(self.ph['fusion_input2'], tf.transpose(self.ph['fusion_input2']))
dist1 = tf.transpose(-2 * d12 + d11) + d11
wt1 = tf.cast((yita[0] * dist1) < tf.reduce_max(dist1), tf.float32)
W2 = tf.matmul(wt1, tf.transpose(wt1))
D2 = tf.matrix_diag(tf.reduce_sum(W2, 1))
L2 = tf.add(W2, -D2)
self.graph_loss2 = beta[1] * tf.trace(tf.matmul(tf.matmul(tf.transpose(self.i_feat), L2), self.i_feat)) / (self.img_batch_size * self.img_batch_size)
d21 = tf.reduce_sum(tf.square(self.ph['fusion_input3']), 1)
d22 = tf.matmul(self.ph['fusion_input3'], tf.transpose(self.ph['fusion_input3']))
dist2 = tf.transpose(-2 * d22 + d21) + d21
wt2 = tf.cast((yita[1] * dist2) < tf.reduce_max(dist2), tf.float32)
W3 = tf.matmul(wt2, tf.transpose(wt2))
D3 = tf.matrix_diag(tf.reduce_sum(W3, 1))
L3 = tf.add(W3, -D3)
self.graph_loss3 = beta[2] * tf.trace(tf.matmul(tf.matmul(tf.transpose(self.t_feat), L3), self.t_feat)) / (self.txt_batch_size * self.txt_batch_size)
self.graph_loss = (self.graph_loss1 + self.graph_loss2 + self.graph_loss3) / 3.0
# classification loss
self.class_loss1 = tf.reduce_mean(tf.nn.l2_loss(self.ph['lab1']-self.class_net1))
self.class_loss2 = tf.reduce_mean(tf.nn.l2_loss(self.ph['lab2']-self.class_net2))
self.class_loss3 = tf.reduce_mean(tf.nn.l2_loss(self.ph['lab3']-self.class_net3))
self.class_loss = alpha * (self.class_loss1 + self.class_loss2 + self.class_loss3) / 3.0
self.acc1 = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(self.ph['lab1'], 1), tf.argmax(self.class_net1, 1)), tf.float32))
self.acc2 = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(self.ph['lab2'], 1), tf.argmax(self.class_net2, 1)), tf.float32))
self.acc3 = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(self.ph['lab3'], 1), tf.argmax(self.class_net3, 1)), tf.float32))
# discrimative loss
pair_domain = tf.concat([tf.ones([self.pair_batch_size, 1]), tf.zeros([self.pair_batch_size, 1])], 1)
img_domain = tf.concat([tf.zeros([self.img_batch_size, 1]), tf.ones([self.img_batch_size, 1])], 1)
txt_domain = tf.concat([tf.zeros([self.txt_batch_size, 1]), tf.ones([self.txt_batch_size, 1])], 1)
domain_label1 = tf.concat([pair_domain, img_domain], 0)
domain_label2 = tf.concat([pair_domain, txt_domain], 0)
domain_loss1 = tf.nn.softmax_cross_entropy_with_logits_v2(logits=self.dis_net1, labels=domain_label1)
domain_loss2 = tf.nn.softmax_cross_entropy_with_logits_v2(logits=self.dis_net2, labels=domain_label2)
self.domain_loss1 = mu * tf.reduce_mean(domain_loss1)
self.domain_loss2 = mu * tf.reduce_mean(domain_loss2)
self.total_loss = self.graph_loss + self.class_loss - self.domain_loss1 - self.domain_loss2
# get variables
self.all_vars = tf.trainable_variables()
self.fn_vars = [v for v in self.all_vars if 'fn_' in v.name]
self.cl_vars = [v for v in self.all_vars if 'cl_' in v.name]
self.dn1_vars = [v for v in self.all_vars if 'dis1_' in v.name]
self.dn2_vars = [v for v in self.all_vars if 'dis2_' in v.name]
def train_model(self):
train_fn = tf.train.AdamOptimizer(learning_rate=self.lr_fn).minimize(self.total_loss, var_list=self.fn_vars+self.cl_vars)
train_dn1 = tf.train.AdamOptimizer(learning_rate=self.lr_dn).minimize(self.domain_loss1, var_list=self.dn1_vars)
train_dn2 = tf.train.AdamOptimizer(learning_rate=self.lr_dn).minimize(self.domain_loss2, var_list=self.dn2_vars)
init = tf.global_variables_initializer()
self.sess.run(init)
for epoch in range(self.TOTAL_EPOCH):
for batch in range(self.BATCH_NUM):
pair_batch = np.hstack((self.I_tr[batch*self.pair_batch_size: (batch+1)*self.pair_batch_size,], self.T_tr[batch*self.pair_batch_size: (batch+1)*self.pair_batch_size,]))
img_batch = np.hstack((self.I_tr[batch*self.img_batch_size: (batch+1)*self.img_batch_size,], np.zeros([self.img_batch_size, self.txt_dim])))
txt_batch = np.hstack((np.zeros([self.txt_batch_size, self.img_dim]), self.T_tr[batch*self.txt_batch_size: (batch+1)*self.txt_batch_size,]))
pair_y_batch = self.L_tr[batch*self.pair_batch_size: (batch+1)*self.pair_batch_size,]
img_y_batch = self.L_tr[batch*self.img_batch_size: (batch+1)*self.img_batch_size,]
txt_y_batch = self.L_tr[batch*self.txt_batch_size: (batch+1)*self.txt_batch_size,]
self.sess.run([train_fn, train_dn1, train_dn2],
feed_dict={self.ph['fusion_input1']: pair_batch,
self.ph['fusion_input2']: img_batch,
self.ph['fusion_input3']: txt_batch,
self.ph['lab1']: pair_y_batch,
self.ph['lab2']: img_y_batch,
self.ph['lab3']: txt_y_batch,
self.ph['kp']: self.kp})
if self.pair:
self.test_paired_model()
else:
self.test_unpaired_model()
def test_paired_model(self):
tr_input = np.hstack((self.I_tr, self.T_tr))
te_img_input = np.hstack((self.I_te, np.zeros(self.T_te.shape)))
te_txt_input = np.hstack((np.zeros(self.I_te.shape), self.T_te))
final_feature = self.sess.run(self.f_feat, feed_dict={self.ph['fusion_input1']: tr_input, self.ph['kp']: 1.0})
txt_feature = self.sess.run(self.i_feat, feed_dict={self.ph['fusion_input2']: te_txt_input, self.ph['kp']: 1.0})
img_feature = self.sess.run(self.t_feat, feed_dict={self.ph['fusion_input3']: te_img_input, self.ph['kp']: 1.0})
Hash = np.sign(final_feature)
H_txt = np.sign(txt_feature)
H_img = np.sign(img_feature)
map_t2i = calc_map(H_txt, Hash, self.L_te, self.L_tr)
map_i2t = calc_map(H_img, Hash, self.L_te, self.L_tr)
print('paired--------------------------------')
print('hash_dim = ' + str(self.hash_dim) + ', alpha = ' + str(alpha) + ', beta = ' + str(beta) + ', mu = ' + str(mu))
print('mapi2t = ' + str(map_i2t) + ', mapt2i = ' + str(map_t2i))
def test_unpaired_model(self):
num_img = self.img_batch_size * self.BATCH_NUM
num_txt = self.txt_batch_size * self.BATCH_NUM
num_fusion = self.pair_batch_size * self.BATCH_NUM
if num_img == 0:
num_img = self.all_num
if num_txt == 0:
num_txt = self.all_num
new_I = np.hstack((self.I_tr[: num_img], np.zeros([num_img, self.txt_dim])))
new_T = np.hstack((np.zeros([num_txt, self.img_dim]), self.T_tr[: num_txt]))
tr_input = np.hstack((self.I_tr, self.T_tr))
te_img_input = np.hstack((self.I_te, np.zeros(self.T_te.shape)))
te_txt_input = np.hstack((np.zeros(self.I_te.shape), self.T_te))
I_L = self.L_tr[: num_img]
T_L = self.L_tr[: num_txt]
new_I_retrieval = self.sess.run(self.i_feat, feed_dict={self.ph['fusion_input2']: new_I, self.ph['kp']: 1.0})
new_T_retrieval = self.sess.run(self.t_feat, feed_dict={self.ph['fusion_input3']: new_T, self.ph['kp']: 1.0})
final_feature = self.sess.run(self.f_feat, feed_dict={self.ph['fusion_input1']: tr_input, self.ph['kp']: 1.0})
txt_feature = self.sess.run(self.i_feat, feed_dict={self.ph['fusion_input2']: te_txt_input, self.ph['kp']: 1.0})
img_feature = self.sess.run(self.t_feat, feed_dict={self.ph['fusion_input3']: te_img_input, self.ph['kp']: 1.0})
Hash = np.sign(final_feature)
H_txt = np.sign(txt_feature)
H_img = np.sign(img_feature)
new_R_I = np.sign(new_I_retrieval)
new_R_T = np.sign(new_T_retrieval)
new_R_I[:num_fusion,] = Hash[:num_fusion,]
new_R_T[:num_fusion,] = Hash[:num_fusion,]
new_map_i2t = calc_map(H_img, new_R_T, self.L_te, T_L)
new_map_t2i = calc_map(H_txt, new_R_I, self.L_te, I_L)
print('unpaired--------------------------------')
print('hash_dim = ' + str(self.hash_dim) + ', alpha = ' + str(alpha) + ', beta = ' + str(beta) + ', mu = ' + str(mu))
print('mapi2t = ' + str(new_map_i2t) + ', mapt2i = ' + str(new_map_t2i))